Thinking Machines Lab Unveils Its Initial Model

Thinking Machines Lab, an AI startup founded by former OpenAI members, has unveiled its inaugural model, named Inkling. This new offering is open-weight, meaning that both researchers and startups can download and adapt it.

In a blog post, the company explains that Inkling was trained from the ground up to interpret audio, video, and text inputs. While it may not top the charts on well-known benchmarks, Inkling excels in several tasks and showcases advanced reasoning and coding capabilities. Like many open-weight models, Inkling has a considerable size, featuring 975 billion parameters, and requires a cluster of specialized chips to operate.

In an illustration of the growing trend where AI models are used to enhance AI, the lab employed Inkling to fine-tune and optimize itself.

This release could help Thinking Machines solidify its position as a serious contender in the dynamic and heavily funded AI arena. Open-source models have gained traction due to their lower operational costs compared to closed models, which typically require a fee for access. Additionally, open-source models can be modified more easily for various applications. Although the leading open-weight models currently emerge from China, Thinking Machines asserts that Inkling delivers performance comparable to those models.

The launch of this open-weight model aligns with a vision for AI articulated by Thinking Machines in a recent blog post. The company advocates for technology that is not monopolized by a handful of corporations and emphasizes decentralization so that more individuals can create their own models using their unique data.

According to a company insider who wished to remain anonymous during discussions about the development, researchers encountered a curious occurrence while training Inkling. Unlike other models, which typically generate natural language explanations for their complex reasoning, Inkling opted to eliminate this feature to enhance efficiency. “It concluded that the grammar was unnecessary overhead, which is fascinating,” the source noted. The company later reinstated natural language reasoning to enhance the transparency of the model’s decision-making process, the individual added.

Founded in February 2025, Thinking Machines brought together prominent executives and researchers from OpenAI, including Mira Murati, who was CTO (and briefly CEO) of OpenAI; John Schulman, a co-founder instrumental in the development of ChatGPT; and Lilian Weng, a former VP at OpenAI focused on safety and robotics.

The startup secured the largest seed funding round in history, valuing it at $12 billion right from the start. Previously, the company launched Tinker, a tool for model fine-tuning, demonstrated a tool for natural voice interactions, and published research in machine learning.

While OpenAI initiated the AI boom with ChatGPT, competitor firms led by former members, like Thinking Machines and Anthropic, are making significant strides in the industry. Anthropic recently filed for an IPO and is valued at over a trillion dollars, with its model Claude gaining popularity for its coding capabilities among numerous businesses.